Publication Details

Download

Michael Chromik, Florian Fincke, Andreas Butz
Mind the (Persuasion) Gap: Contrasting Predictions of Intelligent DSS with User Beliefs to Improve Interpretability
In Companion Proceedings of the 12th ACM SIGCHI Symposium on Engineering Interactive Computing Systems (EICS '20 Companion). Association for Computing Machinery, New York, NY, USA (bib)

Decision support systems (DSS) help users to make more informed and more effective decisions. In recent years, many intelligent DSS (IDSS) in business contexts involve machine learning (ML) methods, which make them inherently hard to explain and comprehend logically. Incomprehensible predictions, however, might violate the users' expectations. While explanations can help with this, prior research also shows that providing explanations in all situations may negatively impact trust and adherence, especially for users experienced in the decision task at hand. We used a human-centered design approach with domain experts to design a DSS for funds management in the construction industry and identified a strong need for control, personal involvement, and adequate data. To create an adequate level of trust and reliance, we contrasted the system's predictions with the values derived from an analytic hierarchical process (AHP), which makes the relative importance of our users' decision-making criteria explicit. We developed a prototype and evaluated its acceptance with 7 construction industry experts. By identifying situations in which the ML prediction and the domain expert potentially disagree, the DSS can identify a persuasion gap and use explanations more selectively. Our evaluation showed promising results and we plan to generalize our approach to a wider range of explainable artificial intelligence (XAI) problems, e.g., to provide explanations with arguments tailored to the user.